SIREN LABS · WEEK 1

What is SIREN Labs measuring this week?

The research bench behind SIREN. Every week we take four or five questions about superintelligence that nobody is measuring well, find open data that can answer them, and build the instrument in public. Each experiment says what it counts and where the numbers come from. None of them is a forecast.

What is on the bench this week?

Is AI building AI yet? · WEEK OF OCT 12, 2026 · 5 EXPERIMENTS

EXPERIMENT 1.1 · The Recursion Meter

How much of the code that builds AI is now written by AI?

BUILD: Count agent-authored pull requests merged into the core AI stacks (PyTorch, Transformers, vLLM, JAX, llama.cpp) each day, as a share of all merged PRs.
DATA: GitHub public API; the agent-PR tracker behind Takeover Watch.
MEASURE: A daily share with a 30-day trend line, recomputable from the PR list.

● ON THE BENCH

EXPERIMENT 1.2 · The Task-Horizon Clock

How long a job can an AI agent finish on its own?

BUILD: Chart the published length of tasks frontier agents complete half the time (METR’s time-horizon measure) and the doubling time it implies.
DATA: METR published results; model release dates.
MEASURE: Hours of human work per task, per model, on a log scale.

● ON THE BENCH

EXPERIMENT 1.3 · Capability Jump Alarm

When does a new model land far off the trend?

BUILD: Fit the trend on public leaderboards and flag any release that beats it by more than two standard deviations.
DATA: LMArena, Epoch AI benchmarks hub, Artificial Analysis.
MEASURE: A list of flagged jumps with the size of the gap, each linked to its source.

● ON THE BENCH

EXPERIMENT 1.4 · The Compute Clock

How fast is frontier training compute growing?

BUILD: A live clock of the largest disclosed training run, with the doubling time from Epoch’s notable-models data.
DATA: Epoch AI notable models dataset (CC BY).
MEASURE: FLOP of the largest known run and months per doubling.

● ON THE BENCH

EXPERIMENT 1.5 · The Lab Silence Index

Do labs go quiet before a big release?

BUILD: Track each lab’s posting cadence (blog, GitHub, X) and mark unusual silences, then check them against what shipped next.
DATA: The newsroom feeds and the Leaders tracker already on SIREN.
MEASURE: Days of silence versus each lab’s normal rhythm, scored after the fact.

● ON THE BENCH

What is coming up on the bench?

WEEK OF OCT 19, 2026Who is watching the watchers?

Safety Team Exodus Tracker
System Card Diff
AGI Timeline Ledger
Chip Export Watch

WEEK OF OCT 26, 2026The physical footprint

Gigawatt Queue
The Open-Weights Gap
Daily Turing Test
Government SI Readiness Map

WEEK OF NOV 2, 2026Brakes and dials

Alignment-to-Capabilities Ratio
AI Incident Clock
Tripwire Registry
All the AGI Odds

What are the lab rules?

Open data onlyEvery number comes from a public source we link to.

Show the methodEach instrument publishes its rules so anyone can recompute it.

Count, don’t predictWe measure what is happening. We don’t say when SI arrives.

Ship or say whyAn experiment either becomes a live room on SIREN or gets a note on why the data wasn’t good enough.

Already shipped from the bench: What AI autonomy signals is SIREN watching? · What if an AI wakes up at 2:14 a.m.? · What AI charts does SIREN publish? · Is ChatGPT down? Is Claude down? · Who makes the most electricity (China vs US)? · What is p(doom)?

Got a question we should measure? How do I report a problem or send an idea to SIREN? Agents can read the plan as JSON.

What questions do people actually ask?

What is SIREN Labs measuring this week?

What is on the bench this week?

This week’s theme is “Is AI building AI yet?”. 5 experiments are on the bench: The Recursion Meter, The Task-Horizon Clock, Capability Jump Alarm, The Compute Clock, The Lab Silence Index.

What is coming up on the bench?

Later weeks stay queued with their themes and experiment names until their start date; nothing ships as a forecast.

What were earlier weeks?

Earlier weeks stay listed with their themes and experiment names so the bench history is checkable.

What are the lab rules?

Open data only (every number from a public source we link to), show the method (each instrument publishes its rules so anyone can recompute it), count don’t predict, and ship or say why (an experiment either becomes a live room on SIREN or gets a note on why the data wasn’t good enough).

Does Labs say when superintelligence arrives?

Count, don’t predict. We measure what is happening. We don’t say when SI arrives.

Is SIREN Labs a doom probability?

No. SIREN Labs builds open-data instruments. The experiments count what is measurable; they are not a probability of harm, and they do not set the SIREN level.

How much of the code that builds AI is now written by AI?

The Recursion Meter: A daily share with a 30-day trend line, recomputable from the PR list.

How long a job can an AI agent finish on its own?

The Task-Horizon Clock: Hours of human work per task, per model, on a log scale.

When does a new model land far off the trend?

Capability Jump Alarm: A list of flagged jumps with the size of the gap, each linked to its source.

How fast is frontier training compute growing?

The Compute Clock: FLOP of the largest known run and months per doubling.

Do labs go quiet before a big release?

The Lab Silence Index: Days of silence versus each lab’s normal rhythm, scored after the fact.

Search-shaped sibling: What AI autonomy signals is SIREN watching? · How do I report a problem or send an idea to SIREN?.